def add_label():
train_path = './dataset/bottle/train_resize/'
train_dataset = pretrain_vgg.MvtecLoader(train_path)
train_loader = DataLoader(train_dataset, batch_size=1, shuffle=False)
# label_name = 'label/bottle_trainK_16.pth'
label_name = 'label/vgg19/bottle/train/16_100.pth'
index_label = torch.load(label_name)
for idx, img in tqdm(train_loader):
idx = idx[0].item()
ironman_grid = plt.GridSpec(1, 1)
fig = plt.figure(figsize=(5,5), dpi=100)
ax = fig.add_subplot(ironman_grid[:,:])
im = ax.imshow(img.squeeze().cpu().detach().numpy().transpose(1,2,0), cmap="Blues")
for i in range(32):
for j in range(32):
label = index_label[idx][i*32+j]
label = label[0]
ax.text(j*8, (i+0.5)*8, label, fontsize=5)
pathToSave = './corn_ground_truth/label10K_' + str(idx) + '.png'
plt.savefig(pathToSave, dpi=100)
plt.close(fig)
def compute_var_black():
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
split_size = 8
train_path = './dataset/bottle/train_resize/'
train_dataset = pretrain_vgg.MvtecLoader(train_path)
train_loader = DataLoader(train_dataset, batch_size=1, shuffle=False)
model = pretrain_vgg.model
model = model.to(device)
white = []
for idx, img in tqdm(train_loader):
idx = idx[0].item()
for y in range(8):
for x in range(8):
patch = img[:, :, 96+x*split_size : 96+x*split_size+split_size, 96+y*split_size : 96+y*split_size+split_size].to(device)
crop_output = model.forward(patch)
crop_output = crop_output.squeeze()
# print(crop_output.size())
white.append(crop_output.cpu().detach().numpy())
print(f'size of white patch: {len(white)}')
# print(white[0].shape)
# print(f'variance of white: {np.var(white, axis=0)}')
draw_hist(np.var(white, axis=0), './var_black.png')
def compute_dis_black():
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
split_size = 8
train_path = './dataset/bottle/train_resize/'
train_dataset = pretrain_vgg.MvtecLoader(train_path)
train_loader = DataLoader(train_dataset, batch_size=1, shuffle=False)
kmeans_path = 'kmeans_bottle_16.pickle'
kmeans = pickle.load(open(kmeans_path, "rb"))
model = pretrain_vgg.model
model = model.to(device)
white = []
white_l1 = []
output_center = kmeans.cluster_centers_[13]
output_center = np.reshape(output_center, (1, -1))
output_center = pil_to_tensor(output_center).to(device)
output_center = torch.squeeze(output_center)
for idx, img in tqdm(train_loader):
idx = idx[0].item()
for y in range(8):
for x in range(8):
patch = img[:, :, 96+x*split_size : 96+x*split_size+split_size, 96+y*split_size : 96+y*split_size+split_size].to(device)
crop_output = model.forward(patch)
crop_output = crop_output.squeeze()
diff = nn.MSELoss()(output_center, crop_output)
white.append(diff.item() ** 0.5)
diff2 = nn.L1Loss()(output_center, crop_output)
white_l1.append(abs(diff2.item()))
# print(crop_output.size())
print(f'size of white patch: {len(white)}')
# print(white[0].shape)
# print(f'mean of white: {np.mean(white)}, variance of white: {np.var(white)}')
# print(f'mean of whiteL1: {np.mean(white_l1)}, variance of whiteL1: {np.var(white_l1)}')
plt.hist(white)
plt.savefig('diff_black.png')
def compute_dis_total():
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
split_size = 8
train_path = './dataset/bottle/train_resize/'
train_dataset = pretrain_vgg.MvtecLoader(train_path)
train_loader = DataLoader(train_dataset, batch_size=1, shuffle=False)
label_name = 'label/bottle_trainK_16.pth'
index_label = torch.tensor(torch.load(label_name))
kmeans_path = 'kmeans/kmeans_bottle_16.pickle'
kmeans = pickle.load(open(kmeans_path, "rb"))
model = pretrain_vgg.model
model = model.to(device)
white = []
white_l1 = []
for idx, img in tqdm(train_loader):
idx = idx[0].item()
for y in range(32):
for x in range(32):
label = index_label[idx][y*32+x].to(device)
label = label.item()
output_center = kmeans.cluster_centers_[label]
output_center = np.reshape(output_center, (1, -1))
output_center = pil_to_tensor(output_center).to(device)
output_center = torch.squeeze(output_center)
patch = img[:, :, x*split_size : x*split_size+split_size, y*split_size : y*split_size+split_size].to(device)
crop_output = model.forward(patch)
crop_output = crop_output.squeeze()
diff = nn.MSELoss()(output_center, crop_output)
white.append(diff.item() ** 0.5)
diff2 = nn.L1Loss()(output_center, crop_output)
white_l1.append(abs(diff2.item()))
# print(crop_output.size())
print(f'size of white patch: {len(white)}')
plt.hist(white)
plt.savefig('diff_total_16_K.png')
def check_training_label():
dataset = pretrain_vgg.MvtecLoader('./dataset/bottle/train_resize/')
loader = DataLoader(dataset, batch_size=1, shuffle=False)
label_name = 'label/vgg19/bottle/train/16_100.pth'
label = torch.load(label_name)
# print(len(label))
# print(len(label[0]))
for idx, img in tqdm(loader):
idx = idx[0].item()
img = np.squeeze(img.detach().cpu().numpy()).transpose((1,2,0))
ironman_grid = plt.GridSpec(1, 1)
fig = plt.figure(figsize=(6,6), dpi=100)
ax1 = fig.add_subplot(ironman_grid[0,0])
im1 = ax1.imshow(img, cmap="Blues")
""" add label text to each patch """
for i in range(32):
for j in range(32):
# print(label[idx][i*32+j])
ax1.text((j+0.2)*8, (i+0.6)*8, label[idx][i*32+j].item(), fontsize=5)
pathToSave = './trainLabelMap/' + str(idx) + '.png'
plt.savefig(pathToSave, dpi=100)
plt.close(fig)